ai-generated-media-pipelinelisted

<!-- last-verified: 2026-07 --> # AI-generated media pipeline Two halves: **generate** the raw asset with an AI model, then **encode it web-ready**. The models churn every few months; the durable value is the encoding, budgeting, and licensing craft in §2–§4 — treat the tool names in §1 as current examples, not the skill. ## 1. Generate — prefer MCP so the agent works in-loop Wire the generator as an **MCP server** where one exists, so an agent can produce and iterate assets inside the build loop instead of round-tripping to a web UI. Discover and vet options with `forge:scout` / `forge:equip`. - **Higgsfield MCP** (cinematic image + video; proxies **Sora 2, Veo 3.1, Kling 3.0, Seedance** for video and **Flux, Seedream, Nano Banana Pro** for stills). Setup: `claude mcp add --transport http --scope user higgsfield https://mcp.higgsfield.ai/mcp` — browser OAuth, no API key. (MCP OAuth won't complete in headless/cron runs — authorize interactively first.) - **Image/still generators**: **Nano Banana / Nano Banana Pro** (Google Gemini image — strong at consistent character/edits and matching an existing frame), Flux, Seedream, Midjourney. - **Video generators** directly: Runway, Luma Dream Machine, Pika, plus the hosted models above. **Prompt for the web, not for a demo reel:** - Generate **at or above** the target display resolution — you can only downscale cleanly. - For a **background loop**, design for seamlessness (first frame ≈ last frame, or reques